Some performance related changes (#16566)

A series of optimizations to reduce allocation rates and speed up
compilation. I think in the end we got 5-10% depending on benchmark. The
optimizations were driven by taking close looks at allocation and cpu
profiles generated by async-profiler.

Author: odersky

community-build/community-projects/requests-scala

@@ -14,24 +14,22 @@ import scala.annotation.internal.sharable
 object Bench extends Driver:
 
   @sharable private var numRuns = 1
-
-  private def ntimes(n: Int)(op: => Reporter): Reporter =
-    (0 until n).foldLeft(emptyReporter)((_, _) => op)
-
+  @sharable private var numCompilers = 1
+  @sharable private var waitAfter = -1
+  @sharable private var curCompiler = 0
   @sharable private var times: Array[Int] = _
 
   override def doCompile(compiler: Compiler, files: List[AbstractFile])(using Context): Reporter =
-    times = new Array[Int](numRuns)
     var reporter: Reporter = emptyReporter
     for i <- 0 until numRuns do
+      val curRun = curCompiler * numRuns + i
       val start = System.nanoTime()
       reporter = super.doCompile(compiler, files)
-      times(i) = ((System.nanoTime - start) / 1000000).toInt
-      println(s"time elapsed: ${times(i)}ms")
-      if ctx.settings.Xprompt.value then
+      times(curRun) = ((System.nanoTime - start) / 1000000).toInt
+      println(s"time elapsed: ${times(curRun)}ms")
+      if ctx.settings.Xprompt.value || waitAfter == curRun + 1 then
         print("hit <return> to continue >")
         System.in.nn.read()
-        println()
     reporter
 
   def extractNumArg(args: Array[String], name: String, default: Int = 1): (Int, Array[String]) = {
@@ -42,20 +40,26 @@ object Bench extends Driver:
 
   def reportTimes() =
     val best = times.sorted
-    val measured = numRuns / 3
+    val measured = numCompilers * numRuns / 3
     val avgBest = best.take(measured).sum / measured
     val avgLast = times.reverse.take(measured).sum / measured
-    println(s"best out of $numRuns runs: ${best(0)}")
+    println(s"best out of ${numCompilers * numRuns} runs: ${best(0)}")
     println(s"average out of best $measured: $avgBest")
     println(s"average out of last $measured: $avgLast")
 
-  override def process(args: Array[String], rootCtx: Context): Reporter =
+  override def process(args: Array[String]): Reporter =
     val (numCompilers, args1) = extractNumArg(args, "#compilers")
     val (numRuns, args2) = extractNumArg(args1, "#runs")
+    val (waitAfter, args3) = extractNumArg(args2, "#wait-after", -1)
+    this.numCompilers = numCompilers
     this.numRuns = numRuns
+    this.waitAfter = waitAfter
+    this.times = new Array[Int](numCompilers * numRuns)
     var reporter: Reporter = emptyReporter
-    for i <- 0 until numCompilers do
-      reporter = super.process(args2, rootCtx)
+    curCompiler = 0
+    while curCompiler < numCompilers do
+      reporter = super.process(args3)
+      curCompiler += 1
     reportTimes()
     reporter
 

compiler/src/dotty/tools/dotc/Bench.scala

@@ -171,7 +171,7 @@ class Driver {
    *  the other overloads without worrying about breaking compatibility
    *  with sbt.
    */
-  final def process(args: Array[String]): Reporter =
+  def process(args: Array[String]): Reporter =
     process(args, null: Reporter | Null, null: interfaces.CompilerCallback | Null)
 
   /** Entry point to the compiler using a custom `Context`.

compiler/src/dotty/tools/dotc/Driver.scala

@@ -231,6 +231,9 @@ class Run(comp: Compiler, ictx: Context) extends ImplicitRunInfo with Constraint
       ctx.settings.Yskip.value, ctx.settings.YstopBefore.value, stopAfter, ctx.settings.Ycheck.value)
     ctx.base.usePhases(phases)
 
+    if ctx.settings.YnoDoubleBindings.value then
+      ctx.base.checkNoDoubleBindings = true
+
     def runPhases(using Context) = {
       var lastPrintedTree: PrintedTree = NoPrintedTree
       val profiler = ctx.profiler

compiler/src/dotty/tools/dotc/Run.scala

@@ -154,14 +154,17 @@ abstract class Positioned(implicit @constructorOnly src: SourceFile) extends Src
     }
   }
 
+  private class LastPosRef:
+    var positioned: Positioned | Null = null
+    var span = NoSpan
+
   /** Check that all positioned items in this tree satisfy the following conditions:
    *  - Parent spans contain child spans
    *  - If item is a non-empty tree, it has a position
    */
   def checkPos(nonOverlapping: Boolean)(using Context): Unit = try {
     import untpd._
-    var lastPositioned: Positioned | Null = null
-    var lastSpan = NoSpan
+    val last = LastPosRef()
     def check(p: Any): Unit = p match {
       case p: Positioned =>
         assert(span contains p.span,
@@ -181,19 +184,19 @@ abstract class Positioned(implicit @constructorOnly src: SourceFile) extends Src
             case _: XMLBlock =>
               // FIXME: Trees generated by the XML parser do not satisfy `checkPos`
             case _: WildcardFunction
-            if lastPositioned.isInstanceOf[ValDef] && !p.isInstanceOf[ValDef] =>
+            if last.positioned.isInstanceOf[ValDef] && !p.isInstanceOf[ValDef] =>
               // ignore transition from last wildcard parameter to body
             case _ =>
-              assert(!lastSpan.exists || !p.span.exists || lastSpan.end <= p.span.start,
+              assert(!last.span.exists || !p.span.exists || last.span.end <= p.span.start,
                 i"""position error, child positions overlap or in wrong order
                    |parent         = $this
-                   |1st child      = $lastPositioned
-                   |1st child span = $lastSpan
+                   |1st child      = ${last.positioned}
+                   |1st child span = ${last.span}
                    |2nd child      = $p
                    |2nd child span = ${p.span}""".stripMargin)
           }
-          lastPositioned = p
-          lastSpan = p.span
+          last.positioned = p
+          last.span = p.span
         p.checkPos(nonOverlapping)
       case m: untpd.Modifiers =>
         m.annotations.foreach(check)

compiler/src/dotty/tools/dotc/ast/Positioned.scala

@@ -830,10 +830,12 @@ trait TypedTreeInfo extends TreeInfo[Type] { self: Trees.Instance[Type] =>
 
   /** The symbols defined locally in a statement list */
   def localSyms(stats: List[Tree])(using Context): List[Symbol] =
-    val locals = new mutable.ListBuffer[Symbol]
-    for stat <- stats do
-      if stat.isDef && stat.symbol.exists then locals += stat.symbol
-    locals.toList
+    if stats.isEmpty then Nil
+    else
+      val locals = new mutable.ListBuffer[Symbol]
+      for stat <- stats do
+        if stat.isDef && stat.symbol.exists then locals += stat.symbol
+      locals.toList
 
   /** If `tree` is a DefTree, the symbol defined by it, otherwise NoSymbol */
   def definedSym(tree: Tree)(using Context): Symbol =

compiler/src/dotty/tools/dotc/ast/TreeInfo.scala

@@ -857,7 +857,7 @@ object tpd extends Trees.Instance[Type] with TypedTreeInfo {
     }
 
     /** After phase `trans`, set the owner of every definition in this tree that was formerly
-     *  owner by `from` to `to`.
+     *  owned by `from` to `to`.
      */
     def changeOwnerAfter(from: Symbol, to: Symbol, trans: DenotTransformer)(using Context): ThisTree =
       if (ctx.phase == trans.next) {
@@ -1144,35 +1144,38 @@ object tpd extends Trees.Instance[Type] with TypedTreeInfo {
       expand(tree, tree.tpe.widen)
   }
 
-  inline val MapRecursionLimit = 10
-
   extension (trees: List[Tree])
 
-    /** A map that expands to a recursive function. It's equivalent to
+    /** Equivalent (but faster) to
      *
      *    flatten(trees.mapConserve(op))
      *
-     *  and falls back to it after `MaxRecursionLimit` recursions.
-     *  Before that it uses a simpler method that uses stackspace
-     *  instead of heap.
-     *  Note `op` is duplicated in the generated code, so it should be
-     *  kept small.
+     *  assuming that `trees` does not contain `Thicket`s to start with.
      */
-    inline def mapInline(inline op: Tree => Tree): List[Tree] =
-      def recur(trees: List[Tree], count: Int): List[Tree] =
-        if count > MapRecursionLimit then
-          // use a slower implementation that avoids stack overflows
-          flatten(trees.mapConserve(op))
-        else trees match
-          case tree :: rest =>
-            val tree1 = op(tree)
-            val rest1 = recur(rest, count + 1)
-            if (tree1 eq tree) && (rest1 eq rest) then trees
-            else tree1 match
-              case Thicket(elems1) => elems1 ::: rest1
-              case _ => tree1 :: rest1
-          case nil => nil
-      recur(trees, 0)
+    inline def flattenedMapConserve(inline f: Tree => Tree): List[Tree] =
+      @tailrec
+      def loop(mapped: ListBuffer[Tree] | Null, unchanged: List[Tree], pending: List[Tree]): List[Tree] =
+        if pending.isEmpty then
+          if mapped == null then unchanged
+          else mapped.prependToList(unchanged)
+        else
+          val head0 = pending.head
+          val head1 = f(head0)
+
+          if head1 eq head0 then
+            loop(mapped, unchanged, pending.tail)
+          else
+            val buf = if mapped == null then new ListBuffer[Tree] else mapped
+            var xc = unchanged
+            while xc ne pending do
+              buf += xc.head
+              xc = xc.tail
+            head1 match
+              case Thicket(elems1) => buf ++= elems1
+              case _ => buf += head1
+            val tail0 = pending.tail
+            loop(buf, tail0, tail0)
+      loop(null, trees, trees)
 
     /** Transform statements while maintaining import contexts and expression contexts
      *  in the same way as Typer does. The code addresses additional concerns:

compiler/src/dotty/tools/dotc/ast/tpd.scala

@@ -83,13 +83,6 @@ object Config {
    */
   inline val failOnInstantiationToNothing = false
 
-  /** Enable noDoubleDef checking if option "-YnoDoubleDefs" is set.
-    * The reason to have an option as well as the present global switch is
-    * that the noDoubleDef checking is done in a hotspot, and we do not
-    * want to incur the overhead of checking an option each time.
-    */
-  inline val checkNoDoubleBindings = true
-
   /** Check positions for consistency after parsing */
   inline val checkPositions = true